Xiaomi MiMo-V2.5-Pro Builds Compiler in 4.3 Hours
Xiaomi unveiled MiMo-V2.5-Pro, an open-weight mixture-of-experts model featuring 1.02 trillion parameters total and 42 billion active per request. The company designed this version for tasks that last hours and involve thousands of tool calls. Internal tests show it programmed a full compiler in 4.3 hours, placing it near Anthropic's Claude Opus 4.6 on coding benchmarks. It also consumes far fewer tokens than Western rivals.
The model processes up to one million tokens in its main version, while the base without retraining reaches 256,000 tokens. Separate encoders convert audio, images, and text into a format for the language model backbone.
Compiler Project in One Afternoon
Xiaomi demonstrated major improvements over the prior version with three examples. First, the team tasked MiMo-V2.5-Pro with creating a complete compiler from a Peking University course, a job that usually takes computer science students weeks.
The model completed the compiler in 4.3 hours through four phases and 672 tool calls. It achieved 233 out of 233 on the hidden test suite, raising test coverage from 59 percent on the initial compile to 100 percent. Xiaomi highlighted the method: the model outlined the full pipeline as a framework, then filled each stage sequentially. The first compile passed 137 of 233 tests. A refactoring step caused a regression, which the model identified and corrected independently.
In the second example, MiMo-V2.5-Pro produced a desktop video editor with about 8,000 lines of code from brief prompts. It operated without supervision for 11.5 hours, using roughly 1,870 tool calls.
For the third, Xiaomi connected the model to a circuit simulator via Claude Code to design a voltage regulator. In under one hour, it met all six technical specifications, with four improving by about an order of magnitude over the initial version.
Token Efficiency and Benchmark Scores
Coding benchmarks give it 78.9 on SWE-bench Verified, 57.2 on SWE-Bench Pro, and 68.4 on Terminal-Bench 2.0. On MiMo Coding Bench, it scores 73.7, near Claude Opus 4.6 at 77.1 and ahead of Gemini 3.1 Pro at 67.8. For agent tasks, it earns 1,581 Elo points on GDPVal-AA and 72.9 on tau3-bench.
Long-context performance stands out on OpenAI's GraphWalks benchmark for navigating node graphs. The earlier MiMo-V2-Pro scored zero at one million tokens, but MiMo-V2.5-Pro manages 0.37 on breadth-first searches and 0.62 on parent node queries at that length.
Stay updated
Get the day's AI and automation news in your inbox. No spam, unsubscribe anytime.
The model builds on MiMo-V2-Flash foundations. Xiaomi reports a blend of local and global attention reduces memory for long texts by almost seven times. A parallel token prediction boosts output speed threefold. Pre-training covered 27 trillion tokens, followed by staged expansion to one million token context.
Post-training employs a teacher-student approach. Specialized models optimize separately for math, security, or tool use. A student model then refines from its trials under guidance from the specialists, merging their abilities.
Additional Models in the Release
Alongside the Pro, Xiaomi released three others. MiMo-V2.5 has 310 billion parameters total, 15 billion active per request. It manages text, images, video, and audio with one million token context. Trained on about 48 trillion tokens, it scores 87.7 on Video-MME, matching Gemini 3 Pro per Xiaomi. Open weights appear on Hugging Face.
MiMo-V2.5-TTS includes three variants: preset voices, new voices from text descriptions, and cloning from short clips. Users add control tags like [crying] or [whispers] in text for pronunciation. These run API-only via Xiaomi's platform, free for now.
MiMo-V2.5-ASR, open weights available, handles Chinese and English, including dialects like Wu, Cantonese, and Hokkien. It manages mid-sentence language switches and song lyrics. On Open ASR Leaderboard, it averages 5.73 percent word error rate, outpacing Gemini 3.1 Pro by over 16 points on dialects and Chinese lyrics.
China's Focus on Open-Weight Volume
Xiaomi's MiMo team continues its late 2025 strategy: multiple models released together, mostly open, aimed at autonomous AI agents. Future plans include larger training scales and improved handling of connections beyond single sentences.
Recently, Xiaomi offered MiMo-V2-Pro, MiMo-V2-Omni, and MiMo-V2-TTS. The prior Pro topped OpenRouter rankings briefly as "Hunter Alpha," with users mistaking it for a Deepseek release. Deepseek now has Deepseek V4, the largest open model, priced low. MiMo-V2.5-Pro enters competition among Chinese open-weight makers, shifting toward cost-effective, prolonged independent task handling.
Xiaomi, a major Chinese electronics firm known for smartphones and smart devices since 2010, expands into AI with these releases.